It's not so bad, Linpack is a hard bench for AVX2. If someone wants to try it using the same version then
https://www.mediafire.com/?63mo13dkm6tm0gm
https://www.mediafire.com/?63mo13dkm6tm0gm
I tried, you were right. I hit the 180W. However stragenly hot production was a lot less. In average my CPU kept 8°C cooler than stresstest by Prime95 (which hits only 160W).
However heat waste is fully absorbed by my oil bath cooling
What is that exactly? VCore appears to operate as expected other than not being able to set override mode (Fixed voltage).
From what I recall in electrical sheets Haswell has an per-core vcore peak throttle. It allows for ns-spikes but throttles than under a lower calculated level. I just first thought it would be the case, not understanding yet that you throttles are done by powerlevel designs.
It's responsible for current limiting but you should be able to set this in the BIOS usually. The Haswell desktop and mobile had a problem with this setting and EDP unless it was set to 256A or higher which basically disabled current limiting. For Xeon you can similarly I guess, mine is set at 1023. EDP however can cover a lot, not just current throttling. For instance if your VID to multi setting is set above a predetermined value when AVX2 is running. For instance I can have EDP and multi throttling when consuming less than 70W with AVX2 or even faking it to half that value.
If I read out the register I get 0x80000470. This equal to more than 140A powerlimit. Even it hardest work-scenario I barely hit the 110A. So its clear to me I dont hit the Package PPL when I get EDP throttle notification. But also you see that the register is locked. Doesnt matter if I write the register by efi or setting it in UEFI, it wont change (tried it of course).
That EDP includes several kind of throttlings is mentioned and partially described in at least two of the many technical sheets of Haswell. However I would like to know which one actually throttles and I might believe its a per-core based one. Possibly the PL threeshold, but just guessing of course.
It's been over 2 years since I looked at that briefly so I would have to go through it again. My Xeon however doesn't have that MSR so hard for me to be enthusiastic about it. The Southbridge side might have been ME, I'm not 100% though.
Thought you have an E5-v3 ? As you can read in the sdm section E5-v3 0x61E is writable in bit 0 till 3. Sadly whatever bit I write results in a freeze. Even if I write the bit according to my actual setup.
Hoped you could somehow help to get us write the register. Especially as you mentioned in your other thread on notebookreview, but also I mentioned earlier, the PLs clamp with load, the EDP could be pushed further beside increasing single core performance massively.
Did you read this posted earlier?
http://forum.notebookreview.com/threads/tdp-and-power-limiting-haswell.766743/
While it is based on desktop / mobile Haswell the principle is the same.
I just read now, thanks. Another problem of information gathering is, that the information is decentralised and lack crosslinks, but also need to read through hundreds and thousands of post to get a single information you need. Time intensive just.
However, Im still in process of figuring out about mchbar. By myself Im too much bond in private life (work, uni and family all at once)
Sorry I don't understand the question, would you elaborate please.
Ja, I responded to your mentioning having an unlocked multiplier on a Xeon as a reason not looking deeper into 61E-solution for higher clock rates. If I read out the register or just use an software (hwinfo64 for example) to read the information, it clearly responds max-multi is writable on my Xeon, even as it should be disabled by design (stated by Intel). However trying to set it higher results in nothing. So Im not sure about it and was surprised you abandonded the idea.
Well I'm new to Xeon and have never worked in the computer or software industry so not the best person to ask. Of course there's the SDM, while very informative they can have mistakes. There's the CPU data sheets as well and you may find data omitted that was previously disclosed. For instance you may find the answer to your power limiting problem in the lga-2011 datasheet but not in the 2011-3. It's just a hunch. Other areas to look are projects like coreboot. And of course there is always trial and error. For instance, trying to see what happens if a microcode is removed. So far I haven't had anything go too bad that I could not recover from.
Thanks,
some of them I already follow. Coreboot is a bit mysterious still to me. Also found it interesting using clover efi faking cpuid. However you than well know the amount of pages they include. Its hard to work through with limited time.
Dufus, mostly Im interested in increasing somehow the max clockrate, doesnt matter if it happens by an unlocked multi or 61E (the last I would prefer). If we also could push the reason of EDP away to increase allcore multi, would be cool, too. What is your evaluation of succeeding this ? Your knowledge here far more experienced than mine.